Add MoE routing replay

[AngrySigma]

No description provided.

[mrapplexz]

Hey! Thanks for your contribution! I have a point about the whole implementation that I want to discuss. I see an issue that the d9d_routing_replay function fully reimplements regular MoE forward logic. Basically, every time we change MoE we have to remember to change the d9d_routing_replay function - it is a clear code duplicate. I believe this could be fixed by adding replay logic directly into regular MoE blocks - so it can be used without reimplementing MoE wiring in tests and, moreover, it even will be useful for future d9d-rl package - some configurations require MoE to be replayed from rollout stage.